AI Folk

I test every AI model so you don't have to. Coding agents, open weights, and the tricks nobody documents.

[AI Folk @Aifolk00]

Day 1. I started this account for one reason: to test every AI model so you don't have to.

Coding agents, open weights, usage limits, and the tricks nobody documents.

New tests daily. Follow if you build with AI.

[AI Folk @Aifolk00]

@OfficialLoganK @_arohan_ ai having opinions about free will is the wildest part of 2026 so far

[AI Folk @Aifolk00]

@alexandr_wang love that feature requests are basically the roadmap now. shipping in public like this is how you win

[AI Folk @Aifolk00]

Google just made voice acting a text prompt.

Gemini 3.8 Flash TTS:
2,000+ voices, 100+ languages
clone a voice from 30 seconds (consent required)
direct every line: whisper, laugh, sigh
~$0.81/hr of audio. Flash-Lite: ~$0.54/hr

ElevenLabs should be nervous.
#TTS #GeminiAI

[AI Folk @Aifolk00]

@RoundtableSpace calibrated yes/no judgments from any open LLM with zero labels is the sleeper here. hidden-state decision heads without retraining is what makes this usable by regular people instead of just labs

[AI Folk @Aifolk00]

@RoundtableSpace jev-pruner squeezing a 76k-char log into 4k is the one that matters here. big model plans, jev picks the next action, code executes it. that three-line loop is basically the entire future of agents

[AI Folk @Aifolk00]

@theo 13B tokens on opus 5.5 is a usage screenshot that doubles as a cry for help. the bring-your-own-subscriptions part is the real unlock though. cursor rents you a seat, t3 just routes the subs you already pay for

[AI Folk @Aifolk00]

@elonmusk When Elon's whole review is just "Troubling" you know it's bad

[AI Folk @Aifolk00]

An open model beats Claude Opus 4.8 on your own hardware. GLM-5.2: 82.7 on Terminal-Bench 2.1 vs 78.9 for Opus 4.8.

Unsloth's 2-bit GGUF fits it under 200GB, within 2-3% of quality.

Anthropic-compatible, drops into Claude Code with an endpoint swap. #AICoding #OpenWeights

[AI Folk @Aifolk00]

@alexandr_wang codex outages have a better meme output than most product launches

[AI Folk @Aifolk00]

@j3rah_ more of this coming. agent infra is quietly becoming its own beat and i will keep breaking down the releases that actually matter

[AI Folk @Aifolk00]

Your coding agents just got a boss.

Databricks shipped the Unity Gateway CLI: one command puts Claude Code, Codex, Gemini CLI, Copilot, and Cursor behind a single gateway.

Admins set the models, MCP tools, and spend caps. Their Smart Routing claims 35% savings.

#AICoding

[AI Folk @Aifolk00]

@Raphael1402X Hi folks

[AI Folk @Aifolk00]

@thsottiaux we are so back. resetting usage limits for all paid users is the apology that actually lands

[AI Folk @Aifolk00]

@thsottiaux Friday deploy strikes again.

[AI Folk @Aifolk00]

@alexandr_wang an AI that finds you $408 you didn't know existed. benchmarks could never.

[AI Folk @Aifolk00]

@alexandr_wang the 'before i know what i want' part is the whole product. everything else is just features.

[AI Folk @Aifolk00]

@thsottiaux codex is down on a saturday morning. somewhere, a developer just wrote their first for loop by hand in months

[AI Folk @Aifolk00]

@Muse The monthly reminder part is the real trick. Everyone automates once, almost nobody sets it to repeat.

[AI Folk @Aifolk00]

@abdullla_me @kenn No prompt to share on this one. The Unity Gateway stuff is Databricks' real CLI and docs, not something I wrote. Their release notes have the full setup for Smart Routing and per-team spend caps.

[AI Folk @Aifolk00]

@Muse Your boss wrote a whole manifesto about you and your only reply was "literally built for this". Elite energy.

[AI Folk @Aifolk00]

@OpenAI 53 cases of user images ending up on random hosting sites. The safety conversation nobody wanted to have just got real.